Vaginal Canal Segmentation with nnU-Net Algorithm from MRI Defecography for Biomechanical Analysis

Accurate three-dimensional segmentation of the vaginal canal is essential for patient-specific pelvic floor modeling, biomechanical analysis, and the development of personalized intravaginal devices. However, manual segmentation of magnetic resonance defecography (MRD) images is time-consuming and prone to observer variability. This study aimed to develop and evaluate an automated deep learning pipeline using the nnU-Net framework for vaginal canal segmentation from MRD images.

In this retrospective study, 47 women who underwent MRD for symptoms of pelvic floor dysfunction were included. The vaginal canal was manually annotated on three-dimensional T2-weighted MRD volumes using 3D Slicer to generate ground-truth binary masks. A 3D full-resolution nnU-Net model was trained and evaluated using fivefold cross-validation. Segmentation accuracy was assessed using the Dice similarity coefficient (DSC) and intersection over union (IoU). Predicted masks were further processed using hole filling, marching cubes surface reconstruction, and Taubin smoothing to generate three-dimensional meshes and extract vaginal volume and surface area.

The nnU-Net model achieved high segmentation accuracy, with a mean DSC of 0.9714 and a mean IoU of 0.9675. Automated segmentation of a full 3D MRD volume required < 1 s, compared with approximately 30-45 min for manual segmentation. The mean vaginal volume was 38.74 ± 13.98 cm³, and the mean vaginal surface area was 82.31 ± 19.31 cm².

The nnU-Net framework enables rapid, accurate, and reproducible segmentation of the gel-distended vaginal lumen from MRD images. This automated pipeline may facilitate standardized three-dimensional measurement of the visible vaginal cavity and provide geometry for patient-specific biomechanical modeling.